The transform is applied on the fly which is faster and consumes less disk space:

train_ds.set_transform(train_transforms)
test_ds.set_transform(val_transforms)

It is common to apply some data augmentations to an image dataset to make a model more robust against overfitting.